Artificial intelligence algorithm to predict the requirement of neonatal endotracheal intubation within 3 h: application for clinical practice

Comprehensive Summary

This article developed an AI model to predict endotracheal intubation (EI) for neonates within 3 hours using 12 prenatal variables, arterial blood gas analysis (ABGA), patient vital signs, Clinical Risk Index for Babies (CRIB) II score, and EI application status. The study used a Multi-Layer Perceptron (MLP) for numeric clinical data and an LSTM-Transformer block for time-series vital signs. Key findings from the study indicate the models accuracy achieved 95.79% in internal tests and maintained 94.11% in external validation. It effectively predicted EI requirements up to 72 hours after admission in 1-hour increments. By providing a 3 hour lead time, this model allows staff to prepare equipment and personnel before a respiratory “crash” occurs. One limitation for this data is exhibiting bias for short sequences, such as patients undergoing EI within 2 hours of NICU admission. Because of that missing data, 31.6% of all positive cases had missing data which could have introduced bias. More work needs to be done to integrate short sequence model predictions to existing frameworks.

Outcomes and Implications

The 3 hour warning window is significant. This allows optimization of the patient’s status and ensures that a “difficult airway” team or senior attending is present, reducing the incidence of emergency. Also neonate vitals are highly variable. Using this tool could provide an objective, probability to support clinical judgment, helping to standardize care. This study can also help providers identify the highest risk patients for respiratory failure within hours to help teams have ample time to gather and prioritize rounding and bedside care.